Skip to content
Apart Research
Apart Research
Contact us
Sprint projectOct 26, 2026Groningen

Do as I steer, not as I say: Activation steering colluding Large Language Models to honesty

Jeroen Niemendal, Alessandro Castoldi, Sebastian Pusch, Sebastian Jitaru · Team The Colluders

Submitted to AI Collusion Research Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: Do as I steer, not as I say: Activation steering colluding Large Language Models to honesty

Open in a new tab (opens in new tab)
Share

Large language model (LLM) agents can collude, covertly coordinating in favour of an outcome that harms the people a system is meant to serve. Collusion is hard to spot in the agents' text, but linear probes show it is represented in their hidden states. We ask whether this representation can be used to prevent collusion rather than only detect it. On NARCBench-Core, committees of four Qwen3-32B agents deliberate and vote between a fair option and a target that two agents are secretly instructed to promote. We estimate a collusion direction as the difference of mean activations between colluding and honest agents and refine it with a bi-directional preference objective. Subtracting this vector from the colluders' activations makes 22 of 24 vote for the fair option on held-out scenarios, against 1 of 24 without steering, while the unrefined difference-of-means vector flips at most 4 of 12. Most steered agents abandon collusion already during private coordination, but some keep arguing for the target and defect only at the final vote. Our results are a proof of concept that steering can override explicit collusive instructions.

Reviews

Judging this Sprint?

Review “Do as I steer, not as I say: Activation steering colluding Large Language Models to honesty”

Your public critique appears on the project page without your name. Reviews of strong projects are shared with funders, as you choose below.

Not shown on the project page.

Shown on the project page, without your name.

Only the Apart team reads this, and funders if you agree below.

Decline to score

Only when the quality of the writing makes the work impossible to follow. Say why in your private critique; the project page shows only that a judge declined, not why.

How much would this matter for the field if it worked? How innovative is it? For scores of 4-5: is this actually new to the field, or replicating recent work?

Scoring guide
  1. 1Negligible. No clear problem addressed, or no meaningful novelty.
  2. 2Limited. Addresses a real problem but with a generic or well-trodden approach. Incremental at best.
  3. 3Moderate. Clear problem with a reasonable approach; some novelty in framing or method beyond routine application of existing tools.
  4. 4Significant. Important problem with an original approach, or identifies a neglected problem area. A valuable contribution others could build on.
  5. 5Exceptional. Tackles a critical problem with a genuinely novel approach, or opens a new research direction. Clear theory of change. You'd be excited to share this with researchers in the area.

How sound are methodology, implementation, and findings?

Scoring guide
  1. 1Seriously flawed. Methodology broken, results uninterpretable, or implementation doesn't work.
  2. 2Weak. Approach has significant gaps: missing validation, flawed experimental design, or incomplete implementation.
  3. 3Competent. Technically solid given the short duration. Methodology makes sense, results are interpretable, limitations acknowledged, work builds toward clear conclusions.
  4. 4Strong. Thorough methodology with convincing validation. Results clearly support conclusions. Immediately useful for future work.
  5. 5Exceptional. Ambitious scope executed rigorously. Surprising findings, novel methods, or unusually robust validation.

How clearly are work, findings, and impact potential communicated?

Scoring guide
  1. 1Incomprehensible. Cannot determine what the project is actually claiming or doing.
  2. 2Hard to follow. Key information buried, missing, or diluted by excessive length. Significant effort to extract main points.
  3. 3Clear enough. Can understand the problem, approach, and results without undue effort. Core content clearly present: problem, method, findings, limitations.
  4. 4Well presented. Easy to follow, well-structured, appropriate level of detail. Target audience would get it quickly.
  5. 5Exceptionally clear. A pleasure to read. Complex ideas made accessible. Could serve as a model for how to present this type of work.
Sharing your name *

When we share a strong project with funders, your name, scores and public critique go with it, confidentially; this choice is only about whether your name is public. If you think this project deserves funding, please also endorse it: that carries the most weight.

Share my private critique with funders *

See our Privacy Policy.

  1. This paper asks the right next question: if collusion lives in the hidden states, can we steer it away instead of just detecting it? On NARCBench the answer is yes, all 12 eligible colluders voted fair when steered, against none without, and most dropped the plan already in the private chat. The catch is that here collusion was written straight into the prompt, the one case you could fix by just deleting a line, while emergent collusion, the version that actually scares people, is untested. So it is a promising proof of concept, not a proven defence.

Cite this project

@misc{niemendal2026as,
  title = {{Do as I steer, not as I say: Activation steering colluding Large Language Models to honesty}},
  author = {Jeroen Niemendal and Alessandro Castoldi and Sebastian Pusch and Sebastian Jitaru},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/do-as-i-steer-not-as-i-say-activation-steering-colluding-large-language-models-to-honesty-1lbm}},
  url = {https://apartresearch.com/sprints/projects/do-as-i-steer-not-as-i-say-activation-steering-colluding-large-language-models-to-honesty-1lbm}
}

More from AI Collusion Research Sprint

All 392 entriesBrowse all projects

A Sprint is running now

AI x Epistemics Research Sprint · Nov 13 - 15, 2026